Post on 21-Jul-2020
On measuring surface-wave phase velocity from station-station1
cross-correlation of ambient signal2
Lapo Boschi (ETH Zurich), Cornelis Weemstra (Spectraseis, ETH Zurich), Julie3
Verbeke (ETH Zurich), Goran Ekstrom (LDEO), Andrea Zunino (Technical4
University of Denmark), Domenico Giardini (ETH Zurich)5
July 30, 20126
Abstract7
We apply two different algorithms to measure surface-wave phase velocity, as a function of8
frequency, from seismic ambient noise recorded at pairs of stations from a large European9
network. The two methods are based on consistent theoretical formulations, but differ in the10
implementation: one method involves the time-domain cross-correlation of signal recorded11
at different stations; the other is based on frequency-domain cross-correlation, and requires12
finding the zero-crossings of the real part of the cross-correlation spectrum. Furthermore,13
the time-domain method, as implemented here and in the literature, practically involves14
the important approximation that interstation distance be large compared to seismic wave-15
length. In both cases, cross-correlations are ensemble-averaged over a relatively long period16
of time (one year). We verify that the two algorithms give consistent results, and infer that17
phase velocity can be successfully measured through ensemble-averaging of seismic ambient18
noise, further validating earlier studies that had followed either approach. The description of19
our experiment and its results is accompanied by a detailed though simplified derivation of20
ambient-noise theory, writing out explicitly the relationships between the surface-wave Green21
function, ambient-noise cross-correlation, and phase and group velocities.22
1 Introduction23
The ability to observe coherent surface-wave signal from the stacked cross-correlation of24
background noise recorded at different stations is essential to improve our resolution of Earth25
structure via seismic imaging. Surface waves generated by earthquakes are best observed26
at teleseismic distances, where the body- and surface-wave packets are well separated, and,27
owing to different geometrical spreading, surface waves are much more energetic than body28
waves; teleseismic surface waves, however, are dominated by intermediate to long periods29
(& 30s), and their speed of propagation is therefore related to mantle, rather than crustal30
structure [e.g., Boschi and Ekstrom, 2002]. The averaged cross-correlated ambient-noise31
signal is instead observed at periods roughly between 5 and 30 s [e.g., Stehly et al., 2006, 2009],32
complementary to the period range of teleseismic surface waves, and allowing to extend33
imaging resolution upwards into the lithosphere-asthenosphere boundary region and the crust.34
1
As first noted by Shapiro and Campillo [2004], the cross-correlation of seismic ambient35
signal recorded at two different stations approximates the Green function associated with a36
point source acting at one of the stations’ location, and a receiver deployed at the other’s.37
Such empirical Green function can then be analyzed in different ways, with the ultimate goal38
of obtaining information about Earth’s structure at various depths between the two stations.39
Most authors either extract group velocity vg from its envelope [e.g., Shapiro et al., 2005;40
Stehly et al., 2006, 2009], or isolate the phase velocity v [e.g., Lin et al., 2008; Nishida et al.,41
2008; Yao and van der Hilst , 2009; Ekstrom et al., 2009]. Fewer authors [e.g., Tromp et al.,42
2010; Basini et al., 2012] attempt to explain (invert) the entire ambient-noise waveform.43
Both v and vg are useful expressions of shallow Earth properties between seismic source44
and receiver, or, in the present case, between two receivers. To measure vg one must be45
able to identify the peak of the surface-wave envelope. This, as a general rule, is easier than46
isolating the carrying sinusoidal wave (i.e. measuring v) at a given frequency. There are,47
however, several properties of vg that make phase-velocity observations useful and possibly48
preferable: (i) the envelope peak is less precisely defined than the phase of the carrying49
sinusoidal wave; (ii) at least so far as the surface-wave fundamental mode is concerned, vg50
depends on, and is in turn used to image, structure over a narrower and shallower depth range51
than v [e.g., Ritzwoller et al., 2001], so that v is particularly helpful to resolve larger depths;52
(iii) a vg measurement needs to be made over a wider time window than a v measurement,53
and contamination by interfering phases is accordingly more likely.54
While the validity of group-velocity estimates based on seismic ambient noise is widely55
recognized, phase velocity is more elusive. For instance, Yao et al. [2006] have noted an im-56
portant discrepancy between two-station observations of phase velocity obtained from tele-57
seismic vs. ambient signal. The systematic application of a far-field approximation, in the58
theoretical expression used to extract the phase from cross-correlation observations (see eqs.59
(41) and (35) below), results in a π/4 shift with respect to the cross-correlation of ballistic60
signal, which has caused some confusion as noted e.g. by Tsai [2009]. We apply here two61
different approaches to measure inter-station surface-wave phase velocity from one year of62
continuous recording at a dense, large array of European stations, first compiled by Verbeke63
et al. [2012b]. Both methods can be derived from the same basic theoretical formulation64
[Tsai and Moschetti , 2010]. One of them is based on time-domain cross-correlation, and is65
implemented, here and elsewhere, using a far-field approximation of the wavefield equation.66
The other is based on frequency domain cross-correlation, and on finding the roots of the real67
part of the cross-correlation spectrum; it involves no far-field approximation. The consistency68
between the two methods’ results further validates earlier phase-velocity tomography studies69
conducted with either approach [e.g., Lin et al., 2008; Yao and van der Hilst , 2009; Ekstrom70
et al., 2009; Fry et al., 2010; Verbeke et al., 2012b].71
2 Theory72
We study the properties of the cross-correlation Cxy(t, ω), function of time t and frequency73
ω, of ambient surface-wave signal u recorded at two seismic instruments, located at positions74
2
Figure 1: Modified from Tsai [2009]. Stations at x and y are separated by a distance ∆x > 0.
Noise sources are far enough that the azimuth θ of any given source is about the same with
respect to either station.
x and y. By definition75
Cxy(t, ω) =1
2T
∫ T
−Tu(x, τ, ω)u(y, t+ τ, ω)dτ, (1)
with the parameter T defining the size of the window over which the cross-correlation is76
computed in practice. We limit our analysis to sources sufficiently far from both receivers for77
the source-receiver azimuth θ to be approximately the same. If we denote ∆x the distance78
separating the two receivers, it then follows, as illustrated in Fig. 1, that the surface wave of79
frequency ω and phase velocity v(ω) generated by a plane-wave source at azimuth θ hits the80
receiver at y with an approximate delay81
td = ∆x cos(θ)/v(ω) (2)
with respect to the one at x.82
Our treatment follows that of Tsai [2009] and Tsai [2011]; we review the formulation car-83
ried out in those works, confirming the theoretical consistency, and pointing out the practical84
differences between the data-analysis methods that we compare. The mathematical treatment85
leads to complete expressions for cross-correlation (section 2.4), and group, as well as phase-86
velocity of the ambient signal (section 2.5). Like Tsai [2009], we assume, as mentioned, that87
sources of ambient noise are far enough from our station pair for the source-receiver azimuth88
to be approximately the same at the two stations.89
Another important assumption of our and most other formulations of ambient-noise theory90
is that the ambient signal be approximately “diffuse”. In practice, this is not true at any91
3
moment in time, but can be at least partially achieved if the ambient signal recorded over a92
very long time (e.g., one year) is subdivided into shorter (e.g., one-day-long) intervals, which93
are then whitened and (after station-station cross-correlation) stacked [Yang and Ritzwoller ,94
2008; Mulargia, 2012]. This procedure is described in detail by Bensen et al. [2007]; we95
refer to it as “ensemble-average”, rather than time-average, since shorter time intervals can96
be chosen to overlap [e.g., Seats et al., 2012; Weemstra et al., 2012]. Over time, an array97
of seismic stations will record ambient signal generated over a wide range of azimuths and98
distances, and the process of stacking simulates the superposition of simultaneously acting99
sources. Stehly et al. [2006] show that, at least in the period range ∼5-15s, most ambient-noise100
signal is likely to be generated by the interaction between oceans and the solid Earth (i.e.,101
ocean storms), and the source distribution of even the stacked ambient signal is accordingly102
nonuniform. Yet, there are both empirical [Derode et al., 2003] and theoretical [Snieder ,103
2004] indications that as long as a significant fraction of ambient signal hits a receiver pair104
along the receiver-receiver azimuth, ensemble-averaging will result in successful applications105
of ambient-noise methods. In our formulation, we treat sources as uniformly distributed in106
azimuth with respect to the receiver pair.107
2.1 Monochromatic signal from a single source108
In the absence of strong lateral heterogeneity in elastic structure, the momentum equation109
for a Love or Rayleigh wave can be decoupled into a differential equation in the vertical, and110
another in the horizontal Cartesian coordinates. The latter coincides with the Helmholtz111
equation and is solved by sinusoidal functions [e.g., Peter et al., 2007].112
Seismic ambient noise can be thought of as the effect of a combination of sources more-or-113
less randomly distributed in space and time. It is however convenient to start our treatment,114
following Tsai [2009], from the simple case of a single source generating a monochromatic115
signal of frequency ω. The first receiver then records a signal116
u(x, t) = S(x, ω) cos(ωt+ φ), (3)
where the constant phase delay φ is proportional to source-receiver distance, and the ampli-117
tude term S(x, ω) is inversely proportional, in the first approximation, to the square-root of118
source-receiver distance (geometrical spreading). The signal (3) is observed at y with a delay119
td, i.e.120
u(y, t) = S(y, ω) cos [ω(t+ td) + φ] . (4)
(By virtue of eq. (2), td is negative when energy propagates from y to x (0 < θ < π/2) and121
positive when energy propagates from x to y.)122
Let us substitute (3) and (4) into (1), so that123
Cxy =S(x)S(y)
2T
∫ T
−Tcos(ωτ + φ) cos [ω(τ + t+ td) + φ] dτ. (5)
It is convenient to substitute z = ωτ , to find124
Cxy =S(x)S(y)
2ωT
∫ ωT
−ωTcos(z + φ) cos [z + φ+ ω(t+ td)] dz. (6)
4
We next make use of the general trigonometric identity cos(A+B) = cosA cosB−sinA sinB,125
valid for any A,B, and126
Cxy =S(x)S(y)
2ωT
∫ ωT
−ωT
{cos2(z) cos [φ+ ω(t+ td)] cos(φ)
+ sin2(z) sin [φ+ ω(t+ td)] sin(φ)
− sin(z) cos(z) cos(φ) sin [φ+ ω(t+ td)]
− sin(z) cos(z) sin(φ) cos [φ+ ω(t+ td)]}
dz,
(7)
which can be simplified if one notices that127 ∫ ωT
−ωTcos2(z)dz =
∫ ωT
−ωT
1 + cos(2z)
2dz = ωT +
1
2sin(2ωT ), (8)
128 ∫ ωT
−ωTsin2(z)dz =
∫ ωT
−ωT
1− cos(2z)
2dz = ωT − 1
2sin(2ωT ), (9)
and finally129 ∫ ωT
−ωTsin(z) cos(z)dz =
[sin2(z)
2
]ωT−ωT
= 0, (10)
where the notation [f(z)]BA = f(B)− f(A).130
After substituting the expressions (8), (9) and (10) into eq. (7),131
Cxy =S(x)S(y)
2
{[1 +
sin(2ωT )
2ωT
]cos(φ) cos [φ+ ω(t+ td)]
+
[1− sin(2ωT )
2ωT
]sin(φ) sin [φ+ ω(t+ td)]
}.
(11)
It then follows from simple trigonometric identities (cosine of the sum, sine of the sum) that132
Cxy =S(x)S(y)
2
{cos [ω(t+ td)] +
sin(2ωT )
2ωTcos [2φ+ ω(t+ td)]
}. (12)
This expression can be simplified if one considers that the size 2T of the cross-correlation133
window should be large compared to the period of the surface waves in question, i.e. T �134
2π/ω, so that 2ωT � 1. Eq. (12) then reduces to135
Cxy ≈S(x)S(y)
2cos [ω(t+ td)] (13)
(compare with eq. (1) of Tsai [2009]). From eq. (13) we infer that the station-station cross-136
correlation of a “ballistic” signal, i.e. generated by a single source localized in space, and not137
scattered, is only useful if the location of the source is known. It coincides (once amplitude138
is normalized) with the response, at one station, to a sinusoidal source located at the other,139
if and only if the two stations are aligned with the source, i.e. azimuth θ = 0 or θ = π, so140
that td = ±∆x/v.141
2.2 Monochromatic signal from a discrete set of sources142
Recorded seismic ambient noise is believed to be the cumulative effect of numerous localized143
sources, distributed almost randomly all around our pair of recording instruments. The signal144
5
generated by a discrete set of monochromatic sources can be written as a superposition of145
single-source signals, eqs. (3) and (4), resulting in146
u(x, t) =∑i
Si(x, ω) cos(ωt+ φi) (14)
and147
u(y, t) =∑i
Si(y, ω) cos[ω(t+ tid) + φi
], (15)
where the summation is over the sources, φi is the phase delay associated with source i,148
and the time delay tid between stations x and y also changes with source azimuth, hence the149
superscript i. In analogy with sec. 2.1, we next substitute (14) and (15) into (1), and150
Cxy =1
2T
∑i,k
{Si(x)Sk(y)
∫ T
−Tcos(ωτ + φi) cos
[ω(τ + t+ tkd) + φk
]dτ
}. (16)
Let us consider the “cross-terms” (cross-correlations between cos(ωτ+φi) and cos[ω(τ + t+ tkd) + φk
]151
with i 6= k) in eq. (16): they are sinusoidal with the same frequency ω but randomly out152
of phase, and therefore do not interfere constructively. The remaining (i = k) terms, on the153
other hand, interfere constructively, as we shall illustrate below, so that, after the contribu-154
tion of a sufficient number of sources has been taken into account, the cross-term contribution155
becomes negligible relative to them. Following other derivations of noise-correlation prop-156
erties, we thus neglect cross-terms from this point on [e.g., Snieder , 2004; Tsai , 2009]. We157
are left with a sum of integrals of the form (5), which we have proved in sec. 2.1 to be158
approximated by (13), so that159
Cxy ≈∑i
Si(x)Si(y)
2cos[ω(t+ tid)
]. (17)
2.3 Continuous distribution of sources160
Eq. (17) can be further generalized to the case of a continuous distribution of sources,161
Cxy ≈∫ ∆x
v
−∆xv
ρ(td, ω) cos [ω(t+ td)] dtd, (18)
where we have introduced the function ρ(td, ω), describing the density of sources as a function162
of inter-station delay td, or, which is the same (recall eq. (2)), azimuth θ. Integration is163
accordingly over td, and the integration limits correspond, through eq. (2), to the interval of164
possible azimuths, from 0 to π. ρ is also a function of ω, as signal generated by differently165
located sources generally has a different frequency content. To keep the notation compact,166
we have incorporated the continuous version of the source term Si(x, ω)Si(y, ω)/2 from eq.167
(17) in the source density function ρ(td, ω).168
In analogy with earlier formulations of ambient-noise theory, we require the source dis-169
tribution to be uniform with respect to azimuth θ. To find the corresponding (not constant)170
expression of ρ as a function of td, we note that, for azimuthally constant source density, ρ(td)171
multiplied by a positive increment |dtd| must coincide with the corresponding increment |dθ|172
times a constant factor. Formally,173
1
2πg(ω)|dθ| = ρ(td, ω)|dtd|, (19)
6
where g(ω)/2π is the normalized value of uniform azimuthal source density, selected so that174
its integral between 0 and 2π is exactly g(ω). The factor g(ω) serves to remind us that source175
amplitude generally changes with frequency. After replacing |dtd| = ∆x sin(θ)|dθ|/v,176
g(ω)
2π|dθ| = ρ(td, ω)
∆x sin[θ(td)]
v|dθ|, (20)
or177
ρ(td, ω) =v(ω)g(ω)
2π∆x sin[θ(td)], (21)
which is the expression of ρ = ρ(td, ω) corresponding to azimuthally uniform source density.178
2.4 Cross-correlation and Green function179
It is convenient to separate the integral in eq. (18) into two integrals, one over positive, and180
the other over negative td,181
Cxy ≈∫ 0
−∆xv
ρ(td, ω) cos [ω(t+ td)] dtd +
∫ ∆xv
0ρ(td, ω) cos [ω(t+ td)] dtd. (22)
The negative- and positive-time contributions to Cxy are usually referred to as anticausal and182
causal, respectively.183
2.4.1 Positive-time (causal) contribution to the cross-correlation184
Let us first consider the second term (td ≥ 0) at the right-hand side of (22), which, since185
ρ(td, ω) is real (see eq. (21)), can be rewritten186
Ctd>0xy ≈ <
[eiωt
∫ ∆xv
0ρ(td, ω)eiωtddtd
], (23)
where <(. . .) equals the real part of its argument. It is convenient to replace ρ(td, ω) with187
its expression (21), and the integration variable td with θ. By differentiating eq. (2), dtd =188
−∆x sin(θ)dθ/v, while the limits of integration 0, ∆x/v correspond to azimuth θ = π/2, 0,189
respectively, hence, using the symmetry of the cosine,190
Ctd>0xy ≈ <
[g(ω)eiωt
2π
∫ π2
0eiω∆x cos(θ)/vdθ
]. (24)
(Recall that positive td corresponds to azimuth 0 < θ < π/2, while the opposite holds for the191
td ≤ 0 term corresponding to π/2 < θ < π.)192
We next rewrite the integral in terms of Bessel and Struve functions. Let us first consider193
the 0-order Bessel function of the first kind in its integral form194
J0(z) =1
π
∫ π
0cos(z sin(θ))dθ (25)
(eq. (9.1.18) of Abramowitz and Stegun [1964]). The integral from 0 to π in (25) can be195
7
transformed into an integral from 0 to π/2:196
J0(z) =1
π
∫ π
0cos(z sin(θ))dθ
=1
π
[∫ π2
0cos(z sin(θ))dθ +
∫ π
π2
cos(z sin(θ))dθ
]
=1
π
[∫ π2
0cos(z sin(θ))dθ −
∫ 0
π2
cos(z sin(π − θ′))dθ′]
=1
π
[∫ π2
0cos(z sin(θ))dθ +
∫ π2
0cos(z sin(θ′))dθ′
]
=2
π
∫ π2
0cos(z sin(θ))dθ.
(26)
We then replace sin(θ) = cos(θ − π/2) and change the integration variable θ = θ′ + π/2,197
J0(z) =2
π
∫ π2
0cos(z cos
(θ − π
2
))dθ
=2
π
∫ 0
−π2
cos(z cos
(θ′))
dθ′
=2
π
∫ π2
0cos(z cos
(θ′))
dθ′,
(27)
and after substituting z with ω∆x/v,198
J0
(ω∆x
v
)=
2
π
∫ π2
0cos
(ω∆x
vcos (θ)
)dθ
=2
π<[∫ π
2
0eiω∆x cos(θ)/vdθ
].
(28)
The 0-order Struve function also has an integral form199
H0(z) =2
π
∫ π2
0sin(z cos(θ))dθ, (29)
which coincides with eq. (12.1.7) of Abramowitz and Stegun [1964] at order 0 and substituting200
Γ(1/2) =√π, with Γ denoting the Gamma function. We replace, again, z with ω∆x/v, and201
H0
(ω∆x
v
)=
2
π=[∫ π
2
0eiω∆x cos(θ)/vdθ
], (30)
with the operator = mapping complex numbers to their imaginary part. It follows from (28)202
and (30) that203 ∫ π2
0eiω∆x cos(θ)/vdθ =
π
2
[J0
(ω∆x
v
)+ iH0
(ω∆x
v
)], (31)
and substituting into (24):204
Ctd>0xy ≈ <
{g(ω)eiωt
4
[J0
(ω∆x
v
)+ iH0
(ω∆x
v
)]}. (32)
Following Tsai [2009], or all other authors conducting ambient-noise analysis in the time205
domain, we next assume that inter-station distance be much larger than the wavelength of the206
8
signal under consideration, i.e. ω∆x/v � 1. It then follows from eq. (9.2.1) of Abramowitz207
and Stegun [1964] that208
J0
(ω∆x
v
)≈√
2 v
ωπ∆xcos
(ω∆x
v− π
4
), (33)
and from eqs. (12.1.34) and (9.2.2) of Abramowitz and Stegun [1964],209
H0
(ω∆x
v
)≈ Y0
(ω∆x
v
)≈√
2 v
ωπ∆xsin
(ω∆x
v− π
4
), (34)
with Y0 denoting the 0-order Bessel function of the second kind.210
Substituting equations (33) and (34) into (32),211
Ctd>0xy ≈ <
{g(ω)eiωt
√v
8πω∆x
[ei(ω∆x/v−π/4)
]}= g(ω)
√v
8πω∆xcos [ω (∆x/v + t)− π/4] .
(35)
Comparing eq. (35) to (13), we note a phase-shift π/4 between the cross-correlated signal212
generated by a teleseismic event aligned with the two stations, and that obtained from the213
ensemble-averaging of seismic ambient noise. π/4 is nothing but the phase-shift between214
a cosine and a Bessel function, for large values of the argument (i.e., in the far field). In215
our experimental set-up, a cosine describes the two-station cross-correlation of a plane wave216
hitting the receivers from a single azimuth; the Bessel function (and hence the π/4 shift)217
emerges from the combined effect of plane waves coming from all azimuths (i.e. focusing over218
the receiver array).219
2.4.2 Negative-time (anticausal) contribution to the cross-correlation220
An analogous treatment applies to the negative-time cross-correlation Ctd<0xy , i.e. the first221
term at the right hand side of eq. (22), which after the variable change from td to θ becomes222
Ctd<0xy ≈ <
[g(ω)eiωt
2π
∫ π
π2
eiω∆x cos(θ)/vdθ
]. (36)
To express also this integral in terms of Bessel and Struve functions, we first notice that223 ∫ π
π2
f(cos(θ))dθ =
∫ π2
0f(
cos(θ′ +
π
2
))dθ′
=
∫ π2
0f(
cos(θ′) cos(π
2
)− sin(θ′) sin
(π2
))dθ′
=
∫ π2
0f(− sin(θ′))dθ′,
(37)
for an arbitrary function f . From eq. (36) it then follows that224
Ctd<0xy ≈ <
[g(ω)eiωt
2π
∫ π2
0e−iω∆x sin(θ)/vdθ
]. (38)
9
Similar to eq. (27) in section 2.4.1, we next replace cos(θ) = sin(θ + π/2) in expression (29)225
for the Struve function, and change the integration variable θ′ = θ + π2 ,226
H0(z) =2
π
∫ π2
0sin (z cos (θ)) dθ
=2
π
∫ π2
0sin(z sin
(θ +
π
2
))dθ
=2
π
∫ π
π2
sin(z sin
(θ′))
dθ′
= − 2
π
∫ π2
0sin(z sin
(θ′))
dθ′.
(39)
Making use of eq. (39), and of expression (26) for the Bessel function J0, with z = ω∆x/v,227
in (38),228
Ctd<0xy ≈ <
{g(ω)e−iωt
4
[J0
(ω∆x
v
)− iH0
(ω∆x
v
)]}, (40)
where only the sign of H0 at the right-hand side has changed with respect to eq. (32). We229
conclude that230
Ctd<0xy ≈ g(ω)
√v
8πω∆xcos [ω (−∆x/v + t) + π/4] , (41)
i.e. the negative-time phase-shift is symmetric to the positive-time one, in agreement with231
Tsai [2009].232
Summing Ctd<0xy (eq. (41)) and Ctd>0
xy (eq. (35)) one finds, according to eq. (22), an233
expression for Cxy valid at all, positive and negative times. To verify its validity, we implement234
it numerically and compare it in Fig. 2 to the result of eq. (17) applied to a very large set of235
sources, for the same frequency and inter-station distance. Confirming earlier findings, the236
two differently computed cross-correlations are practically coincident.237
2.5 Group and phase velocity238
We next consider the more general case of a seismogram formed by the superposition of239
surface waves with different frequencies. Let us start with our expression (35) for the cross-240
correlated signal, grouping the amplitude terms in a generic positive factor S(ω). We then241
find the mathematical expression of a surface-wave packet by (i) discretizing the frequency242
band of interest into a set of closely-spaced frequencies ωi identified by the subscript i, and243
(ii) combining different-frequency contributions by integration around each frequency ωi and244
summation over i, so that245
u(x, t) =∞∑i=1
∫ ωi+ε
ωi−εS(x, ω) cos
[ω
(∆x
v(ω)+ t
)− π
4
]dω, (42)
where ε� ωi. It is convenient to introduce the notation ψ = ω(∆x/v + t)− π/4, and, since246
ε is small, replace it with its Taylor expansion around ωi, i.e.247
ψ(ω) ≈ ψ(ωi) + (ω − ωi)[
dψ
dω
]ωi
, (43)
10
−20 −15 −10 −5 0 5 10 15 20lag [s ]
−1.0
−0.5
0.0
0.5
1.0
Figure 2: Numerical test of expression (41) + (35), with interstation distance of 500 km and
wave speed of 3 km/s. Cxy resulting from the direct implementation of (41) + (35) is denoted
by a solid line. We compare it with the result of applying eq. (17) to model Cxy from the
combined effect of 1000, far, sinusoidal (with 4-s period) out-of-phase sources located at 200
different, uniformly distributed azimuths from the station couple. Finally, we also compute
Cxy from eq. (16) (crosses), neglecting the cross-terms i 6= k; a slight decay, with increasing
lag, in the latter estimate of Cxy is caused by the finite length of the time-integral in the
implementation of (16). Amplitudes have been normalized. All modeled cross-correlations
are perfectly in phase.
where [f(ω)]ωi denotes the value of any function f evaluated at ω = ωi. We rewrite eq. (42)248
accordingly, and find after some algebra that the integral at its right hand side249
∫ ωi+ε
ωi−εS(ω) cos
[ω
(∆x
v(ω)+ t
)− π
4
]dω ≈ S(ωi) cos [ψ(ωi)]
2 sin
{ε[
dψdω
]ωi
}[
dψdω
]ωi
(44)
(valid in the assumption that S be a smooth function of ω). If one introduces a function250
vg(ω) =v(ω)
1− ωv(ω)
dvdω
, (45)
it follows that dψdω takes the compact form251 [
dψ
dω
]ωi
=∆x
vg(ωi)+ t; (46)
we finally substitute it into (44) and substitute the resulting expression into (42), to find252
u(x, t) =
∞∑i=1
S(ωi) cos
[ωi
(∆x
v(ωi)+ t
)− π
4
] 2 sin[ε(
∆xvg(ωi)
+ t)]
[∆x
vg(ωi)+ t] . (47)
Each term at the right-hand side of eq. (47) is the product of a wave of frequency ω and253
speed v(ω) with one of frequency ε� ω and speed vg(ωi). The latter factor, with much lower254
frequency, modulates the signal, and we call “group velocity” its speed vg, which coincides255
with the speed of the envelope of the signal. Eq. (45) shows that, in the absence of dispersion256
(i.e. dvdω = 0) phase and group velocities coincide. In practice, the values of v and vg are257
11
always comparable, and the large difference in frequency results in a large difference in the258
wavelength of the phase and group terms.259
Comparing eq. (47) to (42), it is important to notice that when phase velocity is measured260
from the station-station cross-correlation of ambient signal, a phase correction of π/4 must261
first be applied; the same is not true for group-velocity measurements. We have shown in262
sections 2.4.1 and 2.4.2 that ambient-noise cross-correlation coincides with a combination of263
Bessel functions, and that, for large values of their argument (corresponding to relatively264
large inter-station distance), Bessel functions can be replaced by sinusoidal functions, whose265
argument coincides with the argument of the Bessel functions minus π/4. The π/4-shift in266
(42) and (47) arises precisely from this far-field approximation.267
3 How to measure phase velocity268
To evaluate whether phase velocity can be accurately observed in the ensemble-averaged269
cross-correlation of ambient noise, we use two independent approaches to measure it from270
the same data. Consistency of the results is then an indication of their validity. The first271
approach (section 3.1) consists of cross-correlating and stacking the surface-wave signal (∆t-272
long records of ambient signal in our case) to find the empirical Green function (sec. 2.4),273
from which phase velocity can be measured [e.g., sec. 12.6.2 of Udıas, 1999]. If, as is most274
often the case, one works in the far-field approximation, this requires that a π/4 correction be275
applied to the data as explained in section 2.5 (eq. ( 47)). The other approach we consider is276
based on the result of Aki [1957], confirmed by Ekstrom et al. [2009] for the frequency range277
of interest, that the spectrum of the two-station cross-correlation of seismic ambient noise278
should approximately coincide with a 0-order Bessel function of the first kind (section 3.2);279
in this case, no π/4 correction needs to be applied.280
3.1 Time-domain cross-correlation281
The procedure of ensemble-averaging ambient signal is described in detail, e.g., by Bensen282
et al. [2007]; a long (e.g., one year) continuous seismic record is subdivided into shorter ∆t283
intervals. The records are whitened so that the effects of possible ballistic signal (i.e., large284
earthquakes) present in the data are minimized. The cross-correlation between simultaneous285
∆t-long records from different stations is then computed for all available ∆t intervals, and286
the results for each station pair are stacked over the entire year.287
Bensen et al. [2007] measure group velocity from noise cross-correlations, and suggest that288
phase dispersion can be obtained by integration of group dispersion curves. This approach289
however is not sufficient to identify phase velocity uniquely. Meier et al. [2004] provide an290
algorithm to derive phase velocity from the cross-correlation of teleseismic signals recorded291
by stations aligned with the earthquake azimuth. Fry et al. [2010] and Verbeke et al. [2012a]292
show that the algorithm of Meier et al. [2004] can be successfully applied to the ambient293
signal recorded at a regional-scale array of broadband stations. In reference to the study294
of Fry et al. [2010] where it was first introduced, we shall dub this approach FRY. In the295
following we shall analyze a subset of the phase-dispersion database compiled by Verbeke296
et al. [2012a] via their own automated implementation of FRY.297
The phase-velocity measurements of Verbeke et al. [2012a] are limited to the 0.02-0.1 Hz298
12
frequency range, where seismic ambient noise is known to be strong [Stehly et al., 2009], most299
likely as an effect of ocean storms and the coupling between oceans and the solid Earth [Stehly300
et al., 2006]. Frequency is discretized with increments whose length increases with increasing301
frequency (from 0.02 to 0.05 Hz). For each discrete frequency value, ensemble-averaged cross-302
correlations are (i) band-pass filtered around the frequency in question and (ii) windowed in303
the time-domain via a Gaussian window centered around the time of maximum amplitude of304
(filtered) cross-correlation. Causal and anticausal parts are folded together (i.e. stacked after305
reversing the time-dependence of the anticausal one). The resulting time series is Fourier-306
transformed, and its phase is identified as the arctangent of the ratio of the imaginary to real307
part of the Fourier spectrum, as explained by Udıas [1999], section 12.6.1. Based on eq. (35),308
one must sum π/4 to the resulting folded ensemble-averaged cross-correlation phase before309
applying eq. (3) of Meier et al. [2004] (equivalent to eq. (12.56) of Udıas [1999]). Importantly,310
this π/4 shift is specific to ambient-noise cross-correlation, and must not be applied in two-311
station analysis of ballistic surface-wave signal, as shown by eq. (13). Phase velocity is only312
known up to a 2πn “multiple cycle ambiguity”, with n = 0,±1,±2, .... After iterating over313
the entire frequency band, an array of dispersion curves is found, each corresponding to a314
value of n. Verbeke et al. [2012a] compare each curve (for all integer values of n between315
-5 and 5) with phase velocity as predicted by PREM [Dziewonski and Anderson, 1981], and316
pick the one closest to PREM.317
Ensemble-averaged cross-correlations for two Swiss stations (Fig. 3a) are shown in Fig. 3b.318
At long period (compared to interstation distance divided by wave speed) the causal and319
anti-causal parts of the cross-correlation overlap, complicating the time-domain analysis of320
cross-correlation, whose results are shown in Fig. 3c.321
3.2 Frequency-domain cross-correlation and Bessel-function fitting322
A different method, hereafter referred to as “AKI”, to extrapolate phase velocity from the323
ambient signal recorded at two stations is proposed by Ekstrom et al. [2009], based on much324
earlier work by Aki [1957]. The theoretical basis of this method has been recently rederived325
by Nakahara [2004], Yokoi and Margaryan [2008] and Tsai and Moschetti [2010]. As pointed326
out by Ekstrom et al. [2009], this approach does not require that ω∆x/v � 1, i.e. it will327
work for wavelengths comparable to interstation distance.328
According to AKI, ambient signal recorded over a long time (e.g., one year) is, again,329
subdivided into shorter ∆t intervals. Let us call pi(ω) the frequency spectrum associated330
with a ∆t-long record at station i (Fig. 4a, with ∆t=2 hours). After whitening, this is331
multiplied with the simultaneous ∆t-long recording made at another station j (Fig. 4b),332
resulting in the cross-spectrum, or spectrum of the cross-correlation between the two ∆t-long333
records (Fig. 4c). This procedure is repeated for all available ∆t-intervals in the year, which334
are then stacked together, i.e. ensemble-averaged (Fig. 4d). The resulting quantity is usually335
referred to as “coherency”. Based on Aki [1957],336 ⟨<(
pip∗j
|pi| |pj |
)⟩∝ J0
(ω∆x
v(ω)
), (48)
where < ... > denotes ensemble averaging, the left-hand side is precisely what we call co-337
herency, and the superscript ∗ marks the complex conjugate of a complex number. The338
quantities at the right-hand side of (48) are defined as in section 2.4 above, with ∆x distance339
13
(a)
(b)
(c)
Figure 3: Illustration of the FRY method. (a) Locations (triangles) of stations TORNY
and VDL, from the Swiss broadband network. (b) Ensemble-averaged cross-correlation of
continuous signal recorded at TORNY and VDL, filtered over different frequency bands as
indicated; the bottom trace is the “full” waveform. (c) Array of possible phase-velocity
dispersion curves from cross-correlation of the continuous recordings made at TORNY and
VDL; each curve corresponds to a different value of n, identified by the curve colour as
indicated. The black curve, closest to our selected reference model (PREM), is our preferred
one, but observations are only considered valid in the frequency range marked by black
squares.
14
between stations i and j. (The alert reader might notice at this point that the right-hand340
side of eq. (48) is proportional to Cxy: simply sum, according to eq. (22), its positive- and341
negative-time contributions (32) and (40), respectively [Tsai and Moschetti , 2010].) Again342
based on Aki [1957], the ensemble-averaged imaginary part343 ⟨=(
pip∗j
|pi| |pj |
)⟩= 0. (49)
Importantly, both equations (48) and (49) are shown by Aki [1957] to be valid provided344
that the energy of ambient signal is approximately uniform with respect to azimuth. As345
anticipated at the beginning of section 2, this is typically not true at any moment in time,346
but can be achieved, at least to some extent, by ensemble-averaging [Yang and Ritzwoller ,347
2008].348
Eq. (48) can be used to determine phase dispersion. In practice, observed coherency is349
first of all plotted as a function of frequency (i.e., the ensemble-averaged, whitened cross-350
spectrum is plotted). Values ωi (i = 1, 2, 3, ...) of frequency for which coherency is zero are351
identified. If ω = ωi for some i, the argument of (48) must coincide with one of the known352
zeros zn (n = 1, 2, ...) of the Bessel function J0,353
ωi∆x
v(ωi)= zn. (50)
Eq. (50) can be solved for v,354
v(ωi) =ωi∆x
zn, (51)
and we now have an array of possible measurements of phase velocity at the frequency ωi,355
each corresponding to a different value of n. Implementing (51) at all observed values of ωi,356
an array of dispersion curves is found. Much like in the case of FRY (section 3.1), a criterion357
must then be established to select a unique curve.358
Importantly, the observation of ωi on ensemble-averaged cross-spectra like the one of359
Fig. (4d) is complicated by small oscillations that can be attributed to instrumental noise360
or inaccuracies related to nonuniformity in the source distribution. Before identifying ωi,361
we determine the linear combination of cubic splines that best fits (in least-squares sense,362
via the LSQR algorithm of Paige and Saunders [1982]) observed coherency. Splines are363
equally spaced, and spacing must be selected so that “splined” coherency is sufficiently smooth364
(Fig. 4d).365
Equations (48) and (49) are rarely satisfied by seismic ambient noise as observed in the366
real world. At a given time, the wavefield associated with ambient noise is not diffuse.367
The procedure of ensemble-averaging over a long time serves precisely to mimic a diffuse368
wavefield by combining non-diffuse ones. Yet, there are important systematic effects that369
ensemble-averaging does not remove: in Europe, for example, most of the recorded seismic370
noise is generated in the Atlantic Ocean [Stehly et al., 2006, 2009; Verbeke et al., 2012b],371
and the requirement of an azimuthally uniform source distribution is accordingly not met.372
Presumably, scattering partly compensates for that, but the nonzero observed imaginary part373
of the coherency shown e.g. in Fig. 4d indicates that the problem remains [e.g., Cox , 1973].374
The imaginary part should converge to zero if one ensemble-averages not only over time, but375
also over station-pair azimuth [e.g., Weemstra et al., 2012], but then information on lateral376
Earth structure would be lost.377
15
(a) (b)
(c) (d)
Figure 4: Illustration of the AKI approach. (a) Real part of the power spectrum obtained
by Fourier-transforming two hours of ambient recording at station TORNY. (b) Real part
of the power spectrum from the very same two hours, station VDL. (c) Product of the two
spectra real-parts (coinciding with the real part of the spectrum of the cross-correlation of the
two time-domain signals) obtained after whitening both. (d) Results of ensemble-averaging
an entire year of spectra like the one at (c), for the same two stations: blue and red lines
identify real and imaginary parts, respectively; the black solid line is the linear combination
of cubic splines that best-fits the observed real part of the spectrum. The locations of stations
TORNY and VDL are shown in Fig. 3a.
It is practical to focus the analysis on zero crossings, rather than measuring the overall fit378
between J0 and measured coherency. The latter depends on the power spectrum of the noise379
sources, of which we know very little, and can be affected importantly by data processing380
[Ekstrom et al., 2009].381
4 Application to central European data and cross-validation382
of the two methods383
Fig. 5 shows the set of ∼1000 randomly selected station pairs from Verbeke et al. [2012a] that384
we shall analyze here. The corresponding phase-velocity dispersion curves were measured by385
Verbeke et al. [2012a] following the procedure of section 3.1, after subdividing the entire year386
2006 into day-long intervals and ensemble-averaging the resulting day-long cross-correlations.387
16
Figure 5: (A) Subset of European stations (circles) from Verbeke et al. [2012a] that are also
included in our analysis. We only compare phase-velocity measurements associated with
∼1000 station pairs connected by solid lines. (B) Distribution of epicentral-distance values
sampled by the data set at A.
17
(a) (b)
(c) (d)
Figure 6: Selected FRY phase-velocity dispersion measurements (black circles, connected by a
black line) compared with analogous frequency-domain (AKI) measurements (grey triangles),
for three station pairs: (a) AQU and GIUL, in central Italy, only ∼90 km away from each
other, with a North-South azimuth; (b) TORNY and VDL (see Fig. 3a), with interstation
distance of ∼190 km; (c) WTTA in western Austria and ZCCA in northern Italy, ∼330
km to the south. Triangles in panel (d) mark the locations of all six stations considered
here. We have not yet implemented an algorithm for automatic selection of a preferred AKI
dispersion curve, but the FRY curves clearly fit a single branch of AKI datapoints. At low
frequencies, and particularly at shorter epicentral distances, the match is less accurate. At
longer epicentral distances and high frequencies, occasional one-cycle jumps as in (c) occur.
We apply the AKI method of section 3.2 to continuous records associated with the sta-388
tion pairs of Fig. 5. Our implementation was originally designed for reservoir-scale applica-389
tion [Weemstra et al., 2012], but could be applied to our continent-scale array of data after390
only minor modifications. For each station, continuous recording for the entire year 2006391
is subdivided into intervals of ∆t = 2 hours, with a very conservative 75% overlap between392
neighboring intervals to make sure that no coherent signal traveling from station to station393
is neglected [Seats et al., 2012; Weemstra et al., 2012]. This results in as many as 45 spectra394
per day.395
In Fig. 6 we compare our new phase-velocity measurements with those of Verbeke et al.396
[2012a] for three example station pairs. A visual analysis (which we repeated on many more397
pairs) suggests that the two methods provide very similar results.398
To evaluate quantitatively their level of consistency, we first expand FRY dispersion curves399
over a set of cubic splines, and apply spline interpolation to estimate FRY-based phase-400
18
Figure 7: Frequency of observed phase-velocity misfit (AKI values subtracted from FRY
ones) for the total set of ∼1000 analyzed station pairs. The mean is 13 m/s and the standard
deviation is 151 m/s.
velocity values at the exact frequencies (associated with zero-crossings of the Bessel function)401
where AKI measurements are available. We subtract the AKI phase velocities from the FRY402
ones interpolated at the same frequency, selecting at each frequency the AKI data point403
closest to the FRY one (we thus avoid the well known issue of multiple-cycle ambiguity,404
that equally affects both approaches). We count the number of discrepancy observations,405
independent of frequency, falling in each of a set of 50 m/s intervals, and plot the associated406
histogram in Fig. 7. Both mean and standard deviation of the FRY-AKI discrepancy are407
small (13 m/s and 151 m/s, respectively), and we conclude that, in our implementation,408
the two approaches provide consistent results when applied to the data. Outliers exist with409
misfit larger than ±1000 m/s, but they would not be visible in Fig. 7 even after extending410
the horizontal-axis range.411
We next analyze the dependence of FRY-AKI discrepancy on interstation distance, through412
a second histogram (Fig. 8a) where the misfit is averaged within ∼0.3◦ interstation-distance413
bins. In Fig. 8b the misfit is likewise averaged within 2-mHz increments spanning the whole414
frequency range of interest. Fig. 8a shows that FRY has a tendency to give slightly higher415
velocity estimates with respect to AKI; this effect is reversed at very small and very large in-416
terstation distances. The misfit remains low (∼30 m/s or less) at most interstation distances.417
418
Fig. 8b shows clearly that misfit is systematically smaller (.20 m/s) at relatively high419
19
(a)
(b)
Figure 8: FRY-AKI phase-velocity misfit, for the total set of ∼1000 analyzed station pairs,
averaged within (a) ∼0.3◦ interstation-distance bins, and (b) 2-mHz frequency bins.
20
(a) (b)
Figure 9: (a) FRY-AKI phase-velocity misfit, for the total set of ∼1000 analyzed station
pairs, averaged within (a) ∼0.2◦ × 0.04-Hz distance/frequency bins; (b) number of pairs per
distance/frequency bin.
frequencies (&0.04 Hz) than it is at low frequencies of ∼0.02-0.03 Hz. This is expected, as low420
frequency might result in relatively small ω∆x/v, which would deteriorate the performance421
of FRY (but not of AKI) for short interstation distance ∆x: in practice, the causal and422
anticausal parts tend to overlap in the short-∆x time-domain cross-correlations, making it423
difficult to measure phase via the FRY method [e.g., Ekstrom et al., 2009].424
The combined effect of short ∆x and low frequency is perhaps better illustrated in Fig. 9a,425
where both frequency- and ∆x-dependence of misfit are shown in a single, 2-D plot. It426
emerges that, even at low frequency, AKI and FRY are in good agreement for sufficiently427
large interstation distance. Fig. 9b shows that, not surprisingly, sampling is not uniform with428
respect to frequency and ∆x; most seismic-ambient-noise energy in our station array is found429
at frequencies around ∼0.05 Hz, and some of the discrepancy found at both higher and lower430
frequency (see in particular the top right of Fig. 8a) presubmably reflects the difficulty of431
finding coherent signal in the absence of a sufficiently strong ambient wavefield.432
Overall, averaged discrepancies in Figs. 8 and 9 remain .50 m/s, with the exception of433
the lowest frequencies/shortest epicentral distances considered, where averaged values can434
exceed ∼100 m/s. We take this as an indication that the AKI and FRY methods provide435
essentially consistent results, and we infer that such results can be considered reliable.436
5 Conclusions437
With this study we have conducted a detailed review of the theory of ensemble-averaged cross-438
correlation of surface waves generated by seismic ambient noise, as more tersely described by439
Tsai [2009], Tsai and Moschetti [2010] and Tsai [2011]. With our rederivation we attempt440
to focus the reader’s attention on the potential discrepancy between the time-domain and441
frequency-domain approaches in phase-velocity measurements conducted on ambient-noise442
surface waves. The time-domain approach has generally been applied in the far-field approx-443
imation [e.g., Lin et al., 2008; Yao and van der Hilst , 2009; Fry et al., 2010; Verbeke et al.,444
2012b], and we have emphasized how this approximation is inadequate for interstation dis-445
tances comparable to the seismic wavelengths. The frequency-domain approach of Aki [1957]446
21
and Ekstrom et al. [2009] does not suffer from this limitation.447
We have employed our own implementations of the frequency- (AKI) and far-field time-448
domain (FRY) approaches, to measure Rayleigh-wave phase dispersion from a year of seismic449
noise recorded at a dense array of European stations [Verbeke et al., 2012b]. The two ap-450
proaches provide overall consistent results. As shown in Fig. 9, discrepancies are limited451
to the lowest frequencies and shortest epicentral distances, where the far-field approxima-452
tion on which the FRY method relies does not hold. We infer that Rayleigh-wave phase453
velocity can be successfully observed, via ensemble averaging, from continuous recordings of454
seismic ambient noise, at least within the frequency (∼0.03-0.1 Hz) and inter-station distance455
(∼0.5◦-5◦) ranges analyzed here. We further confirm the validity of published phase-velocity456
observations [e.g., Verbeke et al., 2012b] obtained through the time-domain approach.457
6 Acknowledgments458
This study benefitted from our interactions with Michel Campillo, Bill Fry, Edi Kissling,459
Laurent Stehly, Victor Tsai, and Yang Zha. A. Z. wishes to thank Daniele Spallarossa for his460
advice and constant support.461
References462
Abramowitz, M., and I. A. Stegun, Handbook of Mathematical Functions with Formulas,463
Graphs and Mathematical Tables, National Bureau of Standards Applied Mathematics464
Series, 1964.465
Aki, K., Space and time spectra of stationary waves with special reference to microtremors,466
Bull. Earthquake Res. Inst. Univ. Tokyo, 35 , 415–456, 1957.467
Basini, P., T. Nissen-Meyer, L. Boschi, E. Casarotti, J. Verbeke, O. Schenk, and D. Giardini,468
Adjoint ambient-noise tomography of the European lithosphere, submitted to Geochemistry,469
Geophysics, Geosystems, 2012.470
Bensen, G. D., M. H. Ritzwoller, M. P. Barmin, A. L. Levshin, F. Lin, M. P. Moschetti,471
N. M. Shapiro, and Y. Yang, Processing seismic ambient noise data to obtain reliable472
broad-band surface wave dispersion measurements, Geophys. J. Int., 169 , 1239–1260,473
DOI:10.1111/j.1365–246X.2007.03,374.x, 2007.474
Boschi, L., and G. Ekstrom, New images of the Earth’s upper mantle from measurements of475
surface wave phase velocity anomalies, J. Geophys. Res., 107 , doi:10.1029/2000JB000,059,476
2002.477
Cox, H., Line array performance when signal coherence is spatially dependent, J. Acoust.478
Soc. Am., 54 , 1743–1746, DOI:10.1121/1.1914,473, 1973.479
Derode, A., E. Larose, M. Campillo, and M. Fink, How to estimate the Green’s function of a480
heterogeneous medium between two passive sensors? Application to acoustic waves, Appl.481
Phys. Lett., 83 , 3054–3056, DOI:10.1063/1.1617,373, 2003.482
22
Dziewonski, A. M., and D. L. Anderson, Preliminary reference earth model, Phys. Earth483
Planet. Int., 25 , 297–356, 1981.484
Ekstrom, G., G. A. Abers, and S. C. Webb, Determination of surface-wave phase veloci-485
ties across USArray from noise and Aki’s spectral formulation, Geophys. Res. Lett., 36 ,486
doi:10.1029/2009GL039,131, 2009.487
Fry, B., F. Deschamps, E. Kissling, L. Stehly, and D. Giardini, Layered azimuthal anisotropy488
of Rayleigh wave phase velocities in the European Alpine lithosphere inferred from ambient489
noise, Earth Planet. Sci. Lett., 297 , 95–102, DOI:10.1016/j.epsl.2010.06.008, 2010.490
Lin, F.-C., M. P. Moschetti, and M. H. Ritzwoller, Surface wave tomography of the western491
United States from ambient seismic noise: Rayleigh and Love wave phase velocity maps,492
Geophys. J. Int., 173 , 281–298, DOI:10.1111/j.1365–246X.2008.03,720.x, 2008.493
Meier, T., K. Dietrich, B. Stockhert, and H. Harjes, One-dimensional models of shear wave494
velocity for the eastern Mediterranean obtained from the inversion of Rayleigh-wave phase495
velocities and tectonic implications, Geophys. J. Int., 156 , 45–58, DOI:10.1111/j.1365–496
246X.2004.02,121.x, 2004.497
Mulargia, F., The seismic noise wavefield is not diffuse, J. Acoust. Soc. Am., 131 , 2853–2858,498
doi:10.1121/1.3689,551, 2012.499
Nakahara, H., A systematic study of theoretical relations between spatial correlation and500
Green’s function in one-, two- and three-dimensional random scalar wavefields, Geophys.501
J. Int., 167 , 1097–1105, 2004.502
Nishida, K., H. Kawakatsu, and S. Obara, Three-dimensional crustal S wave velocity structure503
in Japan using microseismic data recorded by Hi-net tiltmeters, J. Geophys. Res., 113 ,504
B10,302, DOI:10.1029/2007JB005,395, 2008.505
Paige, C. C., and M. A. Saunders, LSQR - an algorithm for sparse linear-equations and sparse506
least-squares, ACM Transactions On Math. Software, 8 , 43–71, 1982.507
Peter, D., C. Tape, L. Boschi, and J. H. Woodhouse, Surface wave tomography: global508
membrane waves and adjoint methods, Geophys. J. Int., 171 , 1098–1117, 2007.509
Ritzwoller, N. M., M. H. amd Shapiro, A. L. Levshin, and L. G. M., Crustal and upper mantle510
structure beneath Antarctica and surrounding oceans, J. Geophys. Res., 106 , 30,645–511
30,670, 2001.512
Seats, K. J., J. F. Lawrence, and G. A. Prieto, Improved ambient noise correlation513
functions using Welch’s method, Geophys. J. Int., 188 , 513–523, doi:10.1111/j.1365–514
246X.2011.05,263.x, 2012.515
Shapiro, N. M., and M. Campillo, Emergence of broadband Rayleigh waves from correlations516
of the ambient seismic noise, Geophys. Res. Lett., 31 , 2004.517
Shapiro, N. M., M. Campillo, L. Stehly, and M. H. Ritzwoller, High-resolution surface-wave518
tomography from ambient seismic noise, Science, 307 , 1615–1618, 2005.519
23
Snieder, R., Extracting the Green’s function from the correlation of coda waves: A derivation520
based on stationary phase, Phys. Rev. E , 69 , doi:10.1103/PhysRevE.69.046,610, 2004.521
Stehly, L., M. Campillo, and N. M. Shapiro, A study of the seismic noise from its long-range522
correlation properties, J. Geophys. Res., 111 , 2006.523
Stehly, L., B. Fry, M. Campillo, N. M. Shapiro, J. Guilbert, L. Boschi, and D. Giardini,524
Tomography of the Alpine region from observations of seismic ambient noise, Geophys. J.525
Int., 178 , 338–350, 2009.526
Tromp, J., Y. Luo, S. Hanasoge, and D. Peter, Noise cross-correlation sensitivity kernels,527
Geophys. J. Int., 183 , 791–819,DOI:10.1111/j.1365–246X.2010.04,721.x, 2010.528
Tsai, V. C., On establishing the accuracy of noise tomography travel-time measure-529
ments in a realistic medium, Geophys. J. Int., 178 , 1555–1564, doi:10.1111/j.1365–530
246X.2009.04,239.x, 2009.531
Tsai, V. C., Understanding the Amplitudes of Noise Correlation Measurements, J. Geophys.532
Res., 116 , B09,311, doi:10.1029/2011JB008,483, 2011.533
Tsai, V. C., and M. P. Moschetti, An explicit relationship between time-domain noise534
correlation and spatial autocorrelation (SPAC) results, Geophys. J. Int., 182 , 454–460,535
DOI:10.1111/j.1365–246X.2010.04,633.x, 2010.536
Udıas, A., Principles of Seismology , Cambridge University Press, Cambridge, U.K., 1999.537
Verbeke, J., L. Boschi, D. May, F. Deschamps, and E. Kissling, Anisotropic velocity structure538
of the crust and the upper mantle in central Europe from high resolution ambient noise539
database, Geochemistry Geophysics Geosystems, to be submitted , 2012a.540
Verbeke, J., L. Boschi, L. Stehly, E. Kissling, and A. Michelini, High-resolution Rayleigh-541
wave velocity maps of central Europe from a dense ambient-noise data set, Geophys. J.542
Int., 188 , 1173–1187, doi:10.1111/j.1365–246X.2011.05,308.x, 2012b.543
Weemstra, C., L. Boschi, A. Goertz, and B. Artman, Constraining seismic attenuation from544
recordings of ambient noise, Geophysics, submitted , 2012.545
Yang, Y., and M. H. Ritzwoller, Characteristics of ambient seismic noise as a source546
for surface wave tomography, Geochemistry, Geophysics, Geosystems, 9 , Q02,008,547
doi:10.1029/2007GC001,814, 2008.548
Yao, H., and R. D. van der Hilst, Analysis of ambient noise energy distribution and phase549
velocity bias in ambient noise tomography, with application to SE Tibet, Geophys. J. Int.,550
179 , 1113–1132, doi:10.1111/j.1365–246X.2009.04,329.x, 2009.551
Yao, H., R. D. van der Hilst, and M. V. de Hoop, Surface-wave array tomography in SE Tibet552
from ambient seismic noise and two-station analysis - I. Phase velocity maps, Geophys. J.553
Int., 166 , 732–744,DOI:10.1111/j.1365–246X.2006.03,028.x, 2006.554
Yokoi, T., and S. Margaryan, Consistency of the spatial autocorrelation method with seismic555
interferometry and its consequence, Geophysical Prospecting , 56 , 435–451, 2008.556
24